Bayesian Calibration of Stochastic Agent Based Model via Random Forest.

Area of Science:

• Epidemiology
• Computational Biology
• Data Science

Background:

• Agent-based models (ABMs) are valuable for simulating disease outbreaks and interventions.
• ABMs require extensive calibration due to stochasticity and high parameterization, posing computational challenges.
• Accurate calibration is crucial for the predictive performance of epidemiological models.

Purpose of the Study:

• To develop and demonstrate a random forest-based surrogate modeling technique for accelerating ABM evaluation.
• To apply this technique for calibrating the CityCOVID epidemiological model using Markov chain Monte Carlo (MCMC).
• To compare the performance of this novel calibration method against previous approaches.

Main Methods:

• Utilized a random forest surrogate model to approximate the behavior of the CityCOVID ABM.
• Employed dimensionality reduction techniques, including Principal Component Analysis (PCA) and sensitivity analysis.
• Calibrated the model using Markov chain Monte Carlo (MCMC) to match COVID-19 hospitalization and death data for Chicago (March-June 2020).

Main Results:

• The random forest surrogate model significantly accelerated the evaluation of the epidemiological ABM.
• The MCMC calibration successfully matched observed COVID-19 hospitalization and death data.
• The new method demonstrated improved predictive performance compared to previous approximate Bayesian calibration (IMABC) techniques.
• Achieved a notable reduction in computational cost.

Conclusions:

• Random forest surrogate modeling offers an efficient approach to calibrate complex epidemiological agent-based models.
• This technique enhances the feasibility of using detailed ABMs for real-world public health predictions.
• The developed method provides a computationally tractable solution for high-dimensional model calibration in epidemiology.